Open nikolaydubina opened 1 month ago
Thanks for raising the issue! The reason why distribution-meta-reference-gpu is 7GB is because it uses pytorch as base image which is ~6.18 GB, due to CUDA library for GPU inference. For distributions with remote providers, the size is ~2GB. We are also looking for ways to reduce the size of docker images.
This is likely fixed by #349.
how come docker image is 9GB?
this is not model itself, right?
it is odd to have docker image 20x larger than model itself (e.g. 1B/3B INT4)